Hi piotr, I hope you are doing fine. First I'd like to thank you for your contribution and dedication to the open source community and your innovation. I wanted to ask you regardin S.A.M and the geospatial aspect of it or and deeplearning derived application to the classification of image. Are you familiar or have you tried or have any idea if it is possible to classify different types of buildings ,in my case riyadhs which are a specific type of traditionnal housing in Morocco and other regular buildings, the main goal is to do a time series of the evolution of this type of housing through the years. And if it can be done on GEE (google earth engine ) or just regular collab notebook And finaly if it can be exported as a shapefile. Thank you for your time i hope to hear from you soon and sorry for any inconvenience .
great video! is there an existing approach which handles non-complete segmented objects and makes them complete? E.g. segmenting a car, but the back has been covered by another object.
I’m not sure if it works with TIFF, but personally I would not use this model with medical images. You should aim for top prediction quality not speed optimization.
Such a great video. Can you please tell me how to train Fastsam on a custom dataset like Coco or ade20k to also have image classification and detection
Hi, using several rtx series gpu i found out Fast SAM is almost x100 faster than the original one. Sometimes predictions may seem not so accurate but it's worth noting that, as the authors specified on their github page, fast sam was trained only using 2% of the original sam training dataset. I think many will be ok with such tool working realtime rather than waiting another 20-40 seconds inferencing SAM. How's about yolov8 vs yolo-nas comparison?) testing on my personal datasets has showed yolo-nas was not better than v8 so it would be great if you compare them using different tasks
So you touched on a number of important topics. I think many people will be happy with FastSAM replacing SAM. Like I said for simple scenes FastSAM may be just what you need. 20-40 seconds for SAM inference? I didn't experienced that latency. As for yolov8 vs yolo-nas, you meant pretrained COCO models or fine-tuning?
@@Roboflow I just did inference on a bit higher resolution images so original sam turned out to be too slow As for v8 vs nas i think fine-tuned or trained from scratch will be much cleaner since we can't be actually sure about which data authors used to get their "sota" map values)
@@fourier_mourier same goes for any model. You are never sure about the training process. All you know is the score on COCO. Thats actually one of the biggest problem with model benchmarks...
@@Roboflow yeah, so i would be glad if you have time to compare them too: i used some custom datasets of mine and yolo-nas didn't perform any better than v8 despite their claims =/ Maybe rt-detr can be included too - it seems it has now stable training pipeline in ultralytics package
roboflow is my favorite youtube channel :)
That’s what I wanted to hear 🔥
Thank you so much for your hard work! There are all brilliant videos.
Thank you very much for yor work, and contribution to open source👍
My pleasure!
Another excellent video and demo.
Thanks a lot! I really try my best this time.
Hi piotr,
I hope you are doing fine. First I'd like to thank you for your contribution and dedication to the open source community and your innovation.
I wanted to ask you regardin S.A.M and the geospatial aspect of it or and deeplearning derived application to the classification of image.
Are you familiar or have you tried or have any idea if it is possible to classify different types of buildings ,in my case riyadhs which are a specific type of traditionnal housing in Morocco and other regular buildings, the main goal is to do a time series of the evolution of this type of housing through the years.
And if it can be done on GEE (google earth engine ) or just regular collab notebook
And finaly if it can be exported as a shapefile.
Thank you for your time i hope to hear from you soon and sorry for any inconvenience .
great video! is there an existing approach which handles non-complete segmented objects and makes them complete? E.g. segmenting a car, but the back has been covered by another object.
how can we get .json file from this annotated image which will carry the co-ordinates of the polygon mask as text format???
up
Does this work for medical images in TIFF format ?
I’m not sure if it works with TIFF, but personally I would not use this model with medical images. You should aim for top prediction quality not speed optimization.
Whether to support inference using c++
Such a great video. Can you please tell me how to train Fastsam on a custom dataset like Coco or ade20k to also have image classification and detection
Hi, using several rtx series gpu i found out Fast SAM is almost x100 faster than the original one. Sometimes predictions may seem not so accurate but it's worth noting that, as the authors specified on their github page, fast sam was trained only using 2% of the original sam training dataset. I think many will be ok with such tool working realtime rather than waiting another 20-40 seconds inferencing SAM.
How's about yolov8 vs yolo-nas comparison?) testing on my personal datasets has showed yolo-nas was not better than v8 so it would be great if you compare them using different tasks
So you touched on a number of important topics. I think many people will be happy with FastSAM replacing SAM. Like I said for simple scenes FastSAM may be just what you need. 20-40 seconds for SAM inference? I didn't experienced that latency. As for yolov8 vs yolo-nas, you meant pretrained COCO models or fine-tuning?
@@Roboflow
I just did inference on a bit higher resolution images so original sam turned out to be too slow
As for v8 vs nas i think fine-tuned or trained from scratch will be much cleaner since we can't be actually sure about which data authors used to get their "sota" map values)
@@fourier_mourier same goes for any model. You are never sure about the training process. All you know is the score on COCO. Thats actually one of the biggest problem with model benchmarks...
@@Roboflow yeah, so i would be glad if you have time to compare them too: i used some custom datasets of mine and yolo-nas didn't perform any better than v8 despite their claims =/
Maybe rt-detr can be included too - it seems it has now stable training pipeline in ultralytics package
I think using FastSAM to develop a semi-automatic annotation tool is a great idea.
Thank you for this video❤️, will you pls make a video on speed estimation of vehicle on the road?
Any idea when fastSAM might be upgraded to be based on Yolo11n which is about 7% faster than 8n?
Hi. how can i get results after Segment to txt file to train with yolo ?
I wonder if Sam2/fastsam could be used to improve person tracking.
how to use with MPS
What’s MPS? 🤔
Metal Performance Shaders from Apple@@Roboflow
“device = MPS”
Also handle mobilesam!!!
Not sure if I want us to cover every SAM-based model. Curious why you think MobileSAM is worth takings look?